VersaCamVLA
A robot VLA policy framework for configurable cameras. It learns a scene representation independent of the camera set and injects it into a pretrained VLA backbone, so that the model maintains stable performance across different camera counts and poses.
Timeline (last 90 days)
VersaCamVLA builds a unified scene token interface that maps an arbitrary variable set of posed RGB views into fixed-size latent scene tokens.
VersaCamVLA indirectly trains the scene tokens by predicting the visual features of a set of target views, rather than directly supervising their reconstruction of the input images.
VersaCamVLA does not rely on explicit 3D sensing at deployment and does not need to render any new views, and can be stacked onto existing VLA backbones as a plug-in module.
The paper validates on three platforms: the simulation benchmarks RoboTwin and LIBERO, as well as a real robot platform.
VersaCamVLA consistently outperforms prior VLA methods and direct multi-view baselines on multiple tasks, and maintains stable performance under changes in the number of cameras and unseen camera poses.
VersaCamVLA still currently relies on posed RGB view inputs, and the information bottleneck of the scene tokens is also worth further analysis.
All Facts (6)
VersaCamVLA builds a unified scene token interface that maps an arbitrary variable set of posed RGB views into fixed-size latent scene tokens.
50%UnverifiedVersaCamVLA indirectly trains the scene tokens by predicting the visual features of a set of target views, rather than directly supervising their reconstruction of the input images.
50%UnverifiedVersaCamVLA does not rely on explicit 3D sensing at deployment and does not need to render any new views, and can be stacked onto existing VLA backbones as a plug-in module.
50%UnverifiedVersaCamVLA consistently outperforms prior VLA methods and direct multi-view baselines on multiple tasks, and maintains stable performance under changes in the number of cameras and unseen camera poses.
50%UnverifiedVersaCamVLA still currently relies on posed RGB view inputs, and the information bottleneck of the scene tokens is also worth further analysis.
50%UnverifiedThe paper validates on three platforms: the simulation benchmarks RoboTwin and LIBERO, as well as a real robot platform.
50%